autonomous-anti-scam-agent

Inspiration

Online scams have gotten so sophisticated that ordinary users can’t reliably detect them anymore. Deepfake voices, hyper-personalized phishing, and social-engineering scripts move too fast for manual moderation. We wanted an agent that protects people as they communicate — something that doesn’t just classify risk, but intervenes, stores context for future detection, and autonomously replies in safe ways. The inspiration came from watching family members get hit with impersonation scams and realizing: people need a shield that thinks faster than the attackers.

What it does

Detects scammy messages in real-time Uses RedisVL for vector similarity search Uses TRM Labs risk assessment (stub/fallback) Uses Skyflow for secure storage (stub/fallback) Uses Parallel for autonomous RAG-aware reply generation (stub/fallback) Auto-blocks high-risk senders Stores conversation history for RAG

How we built it

We architected the agent as a real-time pipeline:

1. Message Ingestion & Streaming Incoming messages are normalized and pushed through a lightweight streaming layer that allows sub-100ms risk evaluation.
2. Vector Similarity Detection (RedisVL) We embedded incoming messages using a local LLM encoder and stored them in RedisVL. RedisVL gave us fast vector lookups against a curated dataset of known scam patterns. We also used conversational embeddings to compare the ongoing conversation with historical scam clusters.
3. Risk Scoring (TRM Labs Stub / Fallback) A TRM-like risk engine stub assigns a second-layer risk score — simulating checks like: Transaction-risk heuristics Wallet/phone number suspiciousness Pattern-of-life anomaly detection If unavailable, a fallback deterministic model kicks in.
4. Secure Storage (Skyflow Stub / Fallback) We routed PII-sensitive content through a Skyflow-style vault abstraction, ensuring all sensitive fields are tokenized before hitting our logs or RAG store.
5. Autonomous RAG-Aware Reply Generation (Parallel Stub) We built a small orchestrator that: Pulls relevant context from past conversations Performs “safety rewrites” Generates either a user-facing safe response or an auto-block action The Parallel API is stubbed but modeled after autonomous agent patterns.
6. Auto-Blocking If combined risk > threshold, the agent: Blocks sender Sends a “safety action notice” Stops further message propagation
7. Conversation History Storage We built a lightweight journal that stores: raw message sanitized message embeddings final risk classification This powers long-term RAG-based awareness of scammer behavior. ## Challenges we ran into Latency vs. intelligence tradeoff: Real-time detection required ultra-fast embedding+similarity, which clashed with deeper RAG reasoning. Vector drift: Scam patterns mutate quickly; embedding clusters became stale faster than expected. PII-safe RAG: Ensuring no sensitive identifiers leaked into the RAG store required custom tokenization layers. Stub integration complexity: Modeling TRM, Skyflow, and Parallel behaviors realistically without their real APIs was trickier than expected. Autonomous replies: Preventing the agent from hallucinating or engaging scammers was a safety and alignment challenge. ## Accomplishments that we're proud of Built a near real-time detection and mitigation flow end-to-end. Achieved sub-150ms vector search using RedisVL. Designed a privacy-first pipeline where PII is segregated and tokenized. Created a believable multi-provider architecture (TRM / Skyflow / Parallel) without needing full API access. Developed an agent that can “reason” about conversation context, not just classify single messages. Implemented working autonomous blocking and safe-reply logic. ## What we learned Scam detection must be conversational, not message-by-message. RAG systems need strong data hygiene when dealing with sensitive information. Vector search empowers early detection of patterns traditional classifiers miss. Autonomous agents are powerful but require strict guardrails to avoid unsafe behaviors. Real-time systems demand ruthless performance optimization and async-first design. ## What's next for autonomous-anti-scam-agent Full integration with actual TRM Labs, Skyflow, and Parallel APIs Browser and mobile extensions Multi-lingual and cross-platform scam detection On-device mode for privacy-preserving local detection Community-powered scam pattern library “Family protection mode” that monitors elderly or vulnerable users Behavioral analytics to detect long-con scams (multi-day relationship scams) Visualization dashboards to show scam patterns, risk evolution, and blocked entities

Built With

• parallel
• redisvl
• skyflow
• trmlabs